The properties of metallic materials depend significantly on their crystalline micro-structures, as these have a decisive influence on their strength and deformability.
One example of this is shape-memory alloys, which change shape due to temperature-related alterations in their internal crystalline structure. “Creating the proper micro-structure in these special raw materials is a great technical challenge. Checking this in detail using X-ray analysis demands an immense effort,” says Prof Thomas Niendorf, head of the Metallic Materials Department at the University of Kassel in central Germany.
For this reason, the researchers often used methods from X-ray diffractometry, in which a detector captures diffracted X-ray beams and a software tool translates their intensity into what is known as a pole figure. To determine this number, the material sample has to be rotated and tilted until the measurement data can be converted into the desired pole figure. Using this pole figure, researchers can calculate the order and of the crystals within the metal. These measurements can often take several days to complete. “Our specially developed algorithm makes us three times as fast,” says David Meier, information scientist at the Helmholtz Centre in Berlin and the intelligent embedded systems unit at the University of Kassel.
Machine learning has been used to train the algorithm to create a complete reconstruction of the pole figure from a limited sample of the real measured data in just a few hours. This only differs minimally from the original. The researchers generate pole figures with a random order of grains in the metal using a simulation. With these simulated representations, a custom-adapted deep learning architecture learns how to output the complete pole figure from a given section. This “reconstruction network” can reconstruct the missing pieces of a real measured pole figure from a small section.
When the reconstruction is compared with the real, complete measurement results, it becomes evident that the reconstruction network is able to analyse the sample with sufficient accuracy for this specific example. However, to prove statistically that the method developed works in other, real-world scenarios, it needs to be evaluated in follow-up studies using samples made of different materials. Nonetheless, it appears a highly promising prospect that in the future, this combination of measurement technology and AI could support research and development activities around high-performance, durable materials. The project results were published in the journal Scientific Reports: 13, 5410 (2023): doi.org/10.1038/s41598-023-31580-1.